OpenAI has introduced its new “o1” model family, marking a pivotal shift from the widely recognized GPT series. The o1 models — specifically o1-preview and o1-mini — are designed to handle more complex reasoning tasks than their GPT predecessors, excelling in domains like science, healthcare, and technology. Plus and Team members are already able to access o1 models in ChatGPT. Browsing, file, and image uploading will be available soon. Developers with Tier 5 access can already explore the API.
Key Features of o1 Models
The o1 models prioritize reasoning over mere text generation. While GPT models have excelled in text-based tasks, the o1-preview model, for instance, is designed to tackle problems as complex as those faced by PhD students in physics and biology. It has demonstrated the ability to solve 83% of problems on the International Mathematics Olympiad (IMO) qualifying exam—significantly outperforming GPT-4o, which only solved 13% of the problems. This shift makes the o1 model a powerful tool for industries that require high-level problem-solving.
AGI ACHIEVED 🚀 #openai #o1 pic.twitter.com/b1fNiYsqgy
— Mateusz Tylec (@mateusztylec) September 12, 2024
Additionally, o1-preview excels in coding, ranking in the 89th percentile in Codeforces competitions, showcasing its strength in executing multi-step workflows, debugging, and generating accurate code solutions.
o1-mini: Efficient and Cost-Effective
The o1-mini model is a more streamlined version of o1-preview, designed for faster and cheaper execution of reasoning tasks. While less powerful, it offers an 80% lower cost than o1-preview, making it highly suitable for developers and researchers focused on tasks like coding and math. It performs impressively on the IMO math benchmark, scoring 70%, just shy of o1-preview’s 74%, while being far more cost-effective.
Availability and Access
Both models are now available to ChatGPT Plus users, with o1-preview limited to 30 messages per week and o1-mini capped at 50 messages per week. OpenAI is gradually rolling out access to Enterprise, Team, and Edu users, and plans to make o1-mini available to free-tier ChatGPT users in the future.
Safety and Alignment
A key focus for OpenAI with the o1 models is safety. The o1-preview model scored 84 on one of OpenAI’s toughest jailbreaking tests, compared to a score of 22 for GPT-4o, underscoring its improved ability to handle unsafe prompts and avoid generating harmful or inappropriate content. OpenAI has also partnered with the U.S. and U.K. AI Safety Institutes to further develop and test the safety features of its AI systems.
Conclusion
The launch of the o1 models represents a significant advancement in AI capabilities, pushing beyond the limits of GPT models by focusing on reasoning and multi-step workflows. The integration of o1-preview and o1-mini into the OpenAI ecosystem offers more specialized solutions for developers, researchers, and industries where problem-solving and accuracy are paramount.
